Energy Reduction Through Voltage Scaling and Lightweight Checking

As the semiconductor roadmap reaches smaller feature sizes and the end of Dennard Scaling, design goals change, and managing the power envelope often dominates delay minimization. Voltage scaling remains a powerful tool to reduce energy. We find that it results in about 60% geomean energy reduction on top of other common low-energy optimizations with 22nm CMOS technology. However, when voltage is reduced, it becomes easier for noise and particle strikes to upset a node, potentially causing Silent Data Corruption (SDC). The 60% energy reduction, therefore, comes with a significant drop in reliability. Duplication with checking and triple-modular redundancy are traditional approaches used to combat transient errors, but spending 2–3x the energy for redundant computation can diminish or reverse the benefits of voltage scaling. As an alternative, we explore the opportunity to use checking operations that are cheaper than the base computation they are guarding. We devise a classification system for applications and their lightweight checking characteristics. In particular, we identify and evaluate the effectiveness of lightweight checks in a broad set of common tasks in scientific computing and signal processing. We find that the lightweight checks cost only a fraction of the base computation (0-25%) and allow us to recover the reliability losses from voltage scaling. Overall, we show about 50% net energy reduction without compromising reliability compared to operation at the nominal voltage. We use FPGAs (Field-Programmable Gate Arrays) in our work, although the same ideas can be applied to different systems. On top of voltage scaling, we explore other common low-energy techniques for FPGAs: transmission gates, gate boosting, power gating, low-leakage (high-Vth) processes, and dual-V dd architectures. We do not scale voltage for memories, so lower voltages help us reduce logic and interconnect energy, but not memory energy. At lower voltages, memories become dominant, and we get diminishing returns from continuing to scale voltage. To ensure that memories do not become a bottleneck, we also design an energy-robust FPGA memory architecture, which attempts to minimize communication energy due to mismatches between application and architecture. We do this alongside application parallelism tuning. We show our techniques on a wide range of applications, including a large real-time system used for Wide-Area Motion Imaging (WAMI).